Event ·Multi-source

MIT CSAIL用GPT-6 Astra编辑Qwen3.5-0.8B权重,改动0.016%参数使MATH-500提升6倍

MIT CSAIL研究人员提出Weight-Space Programming权重空间编辑方法,让GPT-6 Astra分析Qwen3.5-0.8B的内部参数,找到与重复推理相关的权重方向并做针对性修改。整个过程不走梯度下降、不做训练,只改动了约0.016%的权重参数。在最强示例中,该模型在MATH-500上的成绩从7.9%提升到48.1%。作者称这类似于给神经网络做基因编辑,通过修改少量参数缓解小模型的推理死循环。

Current read

MIT让GPT-6 Astra直接改Qwen3.5-0.8B的权重,不训练只动0.016%参数,MATH-500从7.9%冲到48.1%,思路很像给模型做基因编辑。

11 sources22° AI heatLast updated 2026/11/3 01:03:33

Evidence

11 sources
Related source 1shao__meng
View source
Related source 2The Rundown AI
View source
Related source 6歸藏(guizang.ai)
View source
Related source 7Thomas Wolf
View source
Related source 9Simon Willison’s Weblog
View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.